【作者】 郑泽忠；

【导师】 范东明； 周国清；

【作者基本信息】 西南交通大学 ， 地图制图学与地理信息工程， 2010， 博士

【摘要】 汽车目标检测是智能交通研究领域中信息采集方面一个非常重要的课题,基于航空影像的汽车目标检测也是图像处理研究的一个热点。因为城市的高速发展,汽车数量的急剧增加,造成交通拥堵异常严重。因此,在交通规划、控制与管理方案的制定过程中,如何保证道路交通网络与城市发展相协调,如何保证交通调查资料的全面性与现势性,优化网络结构,实现城市交通网络的合理布局,使交通网络能充分、高效地发挥作用,显得极为重要。基于高分辨率航空影像进行汽车目标检测研究正好可以充分利用高分辨率航空遥感图像的丰富空间信息,为交通管理部门提供必要的汽车流量信息,从而实现真正的、合理化的“智能”交通。本论文在对高分辨率航空影像数据进行数字镶嵌,裁剪得到典型的高速公路影像基础上,采用最大方差法,边缘检测法,模板匹配法,以及灰度数学形态学和二值数学形态学算法结合进行汽车目标检测研究,本论文围绕汽车目标检测所展开的研究工作主要如下：(1)研究了基于阈值分割算法的最大方差(Otsu)法,通过自动确定最佳阈值,将高分辨率航空高速公路影像二值化,结合二值数学形态学开运算操作进行汽车目标检测。实验结果表明,该算法对于背景简单的影像有很高的检测率；但对于背景复杂的影像进行汽车目标检测,正确率偏低。(2)研究了几种典型的边缘检测二值化算法,结合二值数学形态学算法进行汽车目标检测。实验结果表明,基于Robert算子边缘检测二值化图像边缘连续性不如基于Sobel算子,Prewitt算子的边缘检测二值化图像；基于Sobel算子,Prewitt算子的边缘检测二值化图像效果不如Laplace边缘检测结果二值化图像和Canny边缘检测结果二值化图像；Canny算子是所有边缘检测算子中检测效果最好的。但汽车目标检测研究结果表明,对于简单背景,利用Sobel算子和二值数学形态学方法结合,汽车目标检测率最高；对于复杂背景,利用Canny算子或Sobel算子和二值数学形态学方法结合,汽车目标检测效果最好,但复杂背景汽车目标检测成功率很低。(3)研究了基于模板匹配算法高分辨率航空影像高速公路汽车目标检测。实验结果证明,由于航空影像分辨率高,汽车目标细节清晰,因此,模板匹配算法检测汽车目标的关键在于建立各汽车品牌,各汽车车型的模板库。但模板匹配与最大方差法,边缘检测法相比,其计算量巨大；同时,由于汽车品牌众多,各品牌汽车车型也很多,要建立的模板库工作量也很巨大。(4)研究了基于灰度数学形态学和二值数学形态学算法结合,高分辨率航空影像高速公路汽车目标检测。针对复杂背景,提出了高帽变换和开运算结合,通过筛除大地物(暗背景)及小地物,可以检测到亮背景上的汽车目标；利用低帽变换和闭运算结合,并筛除小地物,可以检测到暗背景上的汽车目标；将开运算和闭运算检测得到的汽车目标叠加,并进行“双影”消除。该算法汽车目标检测调和平均值(Fm)达94%以上,能取得良好的汽车目标检测效果。(5)总的来说,灰度数学形态学和二值数学形态学算法结合用于汽车目标检测,与基于最大方差法、边缘检测算法相比,汽车目标检测准确率更高,具有更强的鲁棒性,但程序运行时间略长。灰度数学形态学和二值数学形态学算法与模板匹配算法相比,具有很强的鲁棒性和高效性。更多还原

【Abstract】 Car detection is a very important task of intelligent transportation system (ITS). And it also has drawn broad attention of research community in computer vision for many years. With the fast development of city, cars increased sharply, leading to jamming very heavily. Thus, how to ensure road net harmonize with the development of city, how to ensure the comprehensive and timely investigation data about transportation, optimize the road net, achieve the reasonable distribution of road net, make it play a full and efficient role.High resolution aerial photo is very affluent in spatial information. Car detection can provide the necessary information for the roads transportation planning organization to realize the reasonable and true intelligent transportation.After image mosaicing of high resolution aerial photos and clipping the typical highway, we utilize maximal variance, edge detection, template matching, gray scale mathematical morphology and two-valued mathematical morphology methods to explore car detection. Our mainly job is listed below.(1) We make use of maximal variance to get the optimal threshold to binarize highway image. Then, we utilize two-valued mathematical morphology to detect cars. The experiment shows that the method has high precise rate for highway with simple background but the precise rate is very low for highway with complex background.(2) We utilize five edge detection arithmetic operators to binarize highway image. Then, we make use of two-valued mathematical morphology on binary images to detect cars. The experiment shows that the edge continuity with Robert operator is worse than Sobel and Prewitt operators. The edge continuity with Canny operator is best than the others.But car detection results indicate that edge detection based on Sobel operator has best precise rate than other operators for highway with simple background. Car detection with Sobel and Canny operators has better results for highway with complex background, but the precise rate is very low for highway with complex background.(3) We also utilize template matching to detect cars in highway. But the experiment shows that because the aerial photo has very high resolution, the detail is very clear, to detect cars in highway, we have to establish exhaustive templates of different brands and models. At the same time, template matching has exhaustive program calculation, because the correlation coefficient will cost plenty of hours.(4) Finally, we make use of gray scale mathematical morphology and two-valued mathematical morphology to detect cars in highway.For the light background, after gray scale morphological top-hat filtering and morphological opening on the highway image, computing the global threshold of top-hat image, we utilize the global threshold to convert the opening image to a binary image, by sieving the bigger and smaller ground objects, the cars can be detected from light background. For the dark background, after gray scale morphological bot-hat filtering and morphological closing on the highway image, computing the global threshold of bot-hat image, we utilize the global threshold to convert the closing image to a binary image, by sieving the smaller ground objects, the cars can be detected from dark background.At last, we overlay the car detection results and eliminate the repeated detection results. The experiment shows that the harmonic mean (Fm) is up to 94%. Thus, the method is very robust.Comparing with maximal variance and edge detection methods, gray scale mathematical morphology and two-valued mathematical morphology method can get higher harmonic mean about car detection. Gray scale mathematical morphology and two-valued mathematical morphology method cost a bit more time than the former, but it is more robust. Comparing with template matching, gray scale mathematical morphology and two-valued mathematical morphology method cost less time, and it is also more efficient and robust.更多还原